A new query engine claims it can fetch evidence for AI agents up to 46 times faster than existing setups.
Researchers built DASE, a system that handles multi-attribute filtering, vector search, and relational joins in a single pass instead of stitching those jobs across separate tools. It includes SemJI, a sparse index built for rare near-neighbor matches, plus an execution layer that runs filter-aware nearest-neighbor searches with batched lookups and score thresholds. On science-focused benchmarks, DASE retrieved evidence for multi-step reasoning queries 6 to 46 times faster than combinations of relational databases, reranking tools, and vector databases, at comparable recall. On an e-commerce benchmark, plugging DASE in as a prefilter raised answer quality from 0.67 to 0.80 while cutting the cost of downstream LLM evaluation from $2.42 to $0.54 per the source's figures.
Most agent pipelines today bolt a vector database onto a relational one and hope the plumbing holds. Every extra hop between systems adds latency and cost, and that tax compounds when an agent has to re-query evidence at each step of a multi-step reasoning chain. Folding filtering, joins, and similarity search into one engine attacks a bottleneck that gets worse, not better, as agentic workflows move from demos to things like scanning scientific literature for evidence.
The numbers come from the paper's own benchmarks, not independent testing, and "beats RDBMS and vector-database baselines" is a claim every new database paper makes. The real test is whether DASE's approach gets adopted by the Postgres-plus-pgvector and dedicated vector-database crowd already entrenched in production agent stacks, not whether it wins on a curated scientific-discovery workload.